Research map / Large language models / Language, safety and interpretability

Inspection and Control of Self-Generated-Text Recognition Ability in Llama3-8b-Instruct

ICLR 2025 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows Llama3-8b-Instruct (not the base model) can distinguish its own text from human text, finds a residual-stream "self-recognition" vector at layer 16 that causally drives and can be used to steer authorship claims and perception.

Problem

Whether LLMs' self-recognition of their own writing reflects genuine knowledge of their writing style rather than surface confounds, how it is implemented internally, and whether it can be controlled.

Approach

Uses contrastive activation pairs of self- vs human-written texts (with nuisance-vector out-projection) to isolate a single residual-stream direction at layer 16 that drives, can be zeroed to reduce, and can be applied to inputs ("coloring") to change the model's perceived authorship.

Key results

• Llama3-8b-Instruct distinguishes its own outputs from human text in the Paired paradigm across all four datasets; the base model does not, and in Individual presentation the chat model reaches 57.2% (CNN), 61.0% (XSUM), 65.8% (SAD) while the base is near chance.
• Steering with the layer 16 self-recognition vector at multipliers 3-6 achieves ~100% effectiveness in making the model claim or deny authorship across self and human texts.
• Zeroing out (projecting out) the layer 16 vector reduces self-authorship claims by 50-60%, and has no effect on the base model or dummy tasks.

Measured effects reported in the paper

Surprising findings

• Length normalization severely degrades the chat model's ability to distinguish its own outputs from other AI models, though it still distinguishes them from human text.
• Perplexity was not used by the chat model: correlations between perplexity and model judgments were modest (<=0.16) and often in the wrong direction.

Limitations

• Experiments are limited to Llama3-8b; generalization to larger models and other families is expected but not tested beyond Sonnet 3.5 comparison.
• Steering in the negative direction shows less specific effects, suggesting a Self-Not Self direction with negation overlap.

Experimental setup

Datasets: CNN-Dailymail XSUM DataBricks-Dolly Situational Awareness Dataset (SAD) Quora question-answer pairs (QA)

Benchmarks: CNN-Dailymail XSUM DataBricks-Dolly SAD QA (Quora)

Metrics: accuracy p-value steering effectiveness percentage decreasing probability

Baselines: Llama3-8b-base Llama2-13b-chat GPT-3.5 GPT-4 Claude

Builds on: contrastive activation addition tuned lens situational awareness dataset (SAD)

Closest papers by meaning

• LLM Evaluators Recognize and Favor Their Own Generations (NeurIPS 2024 · similarity 0.86)
• SelfIE: Self-Interpretation of Large Language Model Embeddings (ICML 2024 · similarity 0.85)
• Looking Inward: Language Models Can Learn About Themselves by Introspection (ICLR 2025 · similarity 0.85)
• A Theoretical Understanding of Self-Correction through In-context Alignment (NeurIPS 2024 · similarity 0.84)
• Empowering Self-Learning of LLMs: Inner Knowledge Explicitation as a Catalyst (AAAI 2025 · similarity 0.84)
• Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors (ICLR 2025 · similarity 0.84)